Nowadays, numerous supervised deep learning models have been investigated for fault diagnosis of rolling element bearings. However, labelling the health states of the bearing vibration data is a time-consuming work and dependent on expert experience. To tackle this problem, this paper explores a new unsupervised bearing fault diagnosis method by proposing an adversarial flow-based model. Flow-based model is a type of generative models that is proved to be better than other types in many aspects in the field of image generation. This paper introduces the flow-based model into the field of machinery fault diagnosis for the first time, and designs an appropriate model architecture so as to train the model in unsupervised and adversarial ways. The proposed model contains an autoencoder, a flow-based model, and a classifier. Firstly, the autoencoder maps the vibration data from signal space to latent vector space. Then, the flow-based model aligns the distributions of the latent vectors of different bearing states with specific prior distributions. Finally, the classifier learns the aligned vector features to discriminate between the two distributions. With the help of distinguishable prior distributions and the adversarial training mechanism between the classifier and the flow-based model together with the autoencoder, the bearing data with the same states are clustered into the same areas. The good clustering performance of the proposed model is validated by a bearing datasets containing 10 types of health states.

Fault diagnosis; Flow-based model; Adversarial training; Unsupervised learning; rolling element bearings